--- license: other library_name: custom inference: false pipeline_tag: image-text-to-text tags: - multimodal - vision-language - deepseek-v4 - moonvit - nvfp4 - fp8 - blackwell - sglang base_model: - MJPansa/DeepSeek-V4-Flash-0731-NVFP4 - deepseek-ai/DeepSeek-V4-Flash-0731 - moonshotai/Kimi-K2.6 --- # DeepSeek V4 Flash 0731 Vision (NVFP4) ![DeepSeek V4 Flash with MoonViT](deepseek-vision-improved.gif) **DeepSeek V4 Flash 0731 with sight.** This development checkpoint connects DeepSeek's reasoning and agentic backbone to the MoonViT vision encoder from [Kimi-K2.6](https://huggingface.co/moonshotai/Kimi-K2.6) through WebBrain's trained, routing-aware PatchMerger projector. This is a complete copy of MJPansa's source-preserving 0731 NVFP4 conversion with WebBrain's frozen MoonViT tower, trained PatchMerger projector, routing bridge, and custom SGLang integration added in the same repository. The text checkpoint is mixed precision: routed expert weights are NVFP4 while attention, shared experts, head, MTP, and other ignored paths preserve the source 0731 formats described by `config.json`, `hf_quant_config.json`, and `conversion-receipt.json`. It is not an all-NVFP4 checkpoint. ## Why vision at WebBrain At [WebBrain](https://www.webbrain.one), we build browser agents that need to understand the visual state of the web—not just extracted text. Screenshots, charts, dashboards, rich editors, and the location and appearance of controls are part of real browser work, so vision is a practical product requirement. In our [American–Chinese open-model frontier benchmark](https://www.webbrain.one/blog/american-chinese-open-model-frontier-gap-benchmark), DeepSeek V4 Flash stood out as a strong and cost-efficient model, while the upstream checkpoint remains text-only. This project attaches a MoonViT vision bridge while keeping the language backbone and vision tower frozen. ## Contents | Component | Provenance | Status | |---|---|---| | Text checkpoint | [`MJPansa/DeepSeek-V4-Flash-0731-NVFP4`](https://huggingface.co/MJPansa/DeepSeek-V4-Flash-0731-NVFP4) at [`64d64cd89bc63a66aa46506da89d7821f7491c62`](https://huggingface.co/MJPansa/DeepSeek-V4-Flash-0731-NVFP4/commit/64d64cd89bc63a66aa46506da89d7821f7491c62) | 48 shards; copied server-side without tensor changes | | Conversion source | [`deepseek-ai/DeepSeek-V4-Flash-0731`](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731) at `9e165c30e2704aec5d9d593cce3eebd58bbef1cb` | 0731 DSpark/MTP tensors preserved | | Vision tower | MoonViT-3d extracted from [`moonshotai/Kimi-K2.6`](https://huggingface.co/moonshotai/Kimi-K2.6) at `7eb5002f6aadc958aed6a9177b7ed26bb94011bb` | 329 BF16 tensors; frozen | | Projector | WebBrain PatchMerger, trained with the text backbone and tower frozen | 6 BF16 tensors; 40,119,040 parameters | | Serving | External DeepSeek/MoonViT SGLang model and processor | Experimental; source patch required | The added components are byte-identical to the published WebBrain vision artifacts: | File | Size | SHA-256 | |---|---:|---| | `vision_tower.safetensors` | 833,765,768 bytes | `1382c41f1a4afc91791ade630e2b1e1cef68cc5a1e09668a45970a5d5e1b8f15` | | `mm_projector.safetensors` | 80,238,568 bytes | `7024d9d5c9714c7abbc09abda015f083b7d7b107745eb78879f019bf4721577a` | The initial mirror commit, `3457594532d890f8a46c18dda72d6a85c94ab418`, matched all 78 source paths, sizes, Git blobs, and LFS SHA-256 values (175,573,280,882 bytes). The final package intentionally replaces the model card and augments `config.json` and `model.safetensors.index.json` so the vision components can be discovered. The 48 text shards remain unchanged. See [`VISION_ADAPTER_MANIFEST.json`](VISION_ADAPTER_MANIFEST.json) and the preserved [`conversion-receipt.json`](conversion-receipt.json) for machine-readable provenance. ## Build status - [x] Pin and exactly mirror the source-preserving 0731 NVFP4 conversion. - [x] Preserve all 48 text shards, the conversion receipt, and 0731 DSpark/MTP tensors. - [x] Preserve mixed-precision metadata: routed experts are NVFP4 while ignored attention, shared-expert, head, and MTP paths retain their source formats. - [x] Add and fingerprint-verify the frozen BF16 MoonViT tower. - [x] Add and fingerprint-verify the trained BF16 PatchMerger projector. - [x] Add the routing palette, external processor/model package, source patch, launch wrapper, and regression tests. - [x] Augment the source config and index without changing upstream text tensor mappings or quantization metadata. - [x] Pass a full-model GPU loader/startup gate and live image-generation smoke test with the established deployment path. - [ ] Run fresh text-parity and live image-generation smoke tests on the target Blackwell hardware and SGLang image. ## Integration details The projector performs: ```text LayerNorm -> 2x2 patch merge -> Linear(4608, 4608) -> GELU -> Linear(4608, 4096) ``` Text token routing IDs are preserved. Image positions receive deterministic routing IDs from the checked-in 64-ID palette. The integration currently supports one literal `` marker and targets at most 512 merged image tokens. This is not a stock Transformers or stock SGLang multimodal architecture. Use the checked-in external model/processor package and the SGLang revision pinned in [`docs/SGLANG_DEPLOYMENT.md`](docs/SGLANG_DEPLOYMENT.md). ## Validation status - Exact source mirror metadata was verified before the vision overlay. - Both added safetensors files were verified by LFS SHA-256 after publication. - The source config's architecture, 0731 DSpark fields, mixed FP8/NVFP4 quantization metadata, and per-layer routed-expert map are preserved; only vision metadata was added. - The index still references the same 48 text shards and additionally maps the 335 vision/projector tensor names. - **A full GPU loader/startup and image-generation smoke test has not yet been rerun for this assembled 0731 checkpoint.** Earlier validation of the same vision adapter and SGLang glue with another NVFP4 text package is not claimed as validation here. ## Usage Download the complete repository, then point the included wrapper at that local snapshot: ```bash export DEEPSEEK_VISION_MODEL_PATH=/models/DeepSeek-V4-Flash-0731-Vision-NVFP4 export DEEPSEEK_VISION_PYTHONPATH="$DEEPSEEK_VISION_MODEL_PATH/sglang_ext" export DEEPSEEK_VISION_KERNEL_PROFILE=blackwell-native export DEEPSEEK_VISION_TP=4 scripts/launch_sglang_moonvit.sh ``` Treat the Blackwell profile as a starting configuration, not a validation claim. Begin with SGLang's native `/generate` endpoint as documented in the deployment guide. OpenAI chat image parts, production concurrency, CUDA graphs, and broad quality behavior remain unvalidated. ## Licenses The converted text files retain the terms and notices from their source. The MoonViT-derived tower is covered by `LICENSE_KIMI_K2.6`. The WebBrain projector and integration source are provided under the terms included with this repository. Downstream users are responsible for satisfying all applicable licenses. ## Method credit The overall construction and model-card approach was inspired by [Baseten's GLM-5.2-Vision-NVFP4](https://huggingface.co/baseten/GLM-5.2-Vision-NVFP4): keep the text backbone and MoonViT tower frozen, train a compact PatchMerger projector between them, and publish provenance and hardware constraints explicitly. Credit to the Baseten team for demonstrating this practical recipe. This project adds a DeepSeek-specific routing bridge so mixed text/image embeddings preserve hash-routed text behavior. It does not reuse Baseten model weights, benchmark results, or deployment artifacts. ## Acknowledgements Built on [DeepSeek AI's DeepSeek V4 Flash 0731](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731), [MJPansa's source-preserving 0731 NVFP4 conversion](https://huggingface.co/MJPansa/DeepSeek-V4-Flash-0731-NVFP4), and [Moonshot AI's Kimi-K2.6](https://huggingface.co/moonshotai/Kimi-K2.6), with the vision-attachment method inspired by [Baseten's GLM-5.2-Vision-NVFP4](https://huggingface.co/baseten/GLM-5.2-Vision-NVFP4). These teams were not involved in this experimental development checkpoint; please do not direct issues with this repository to them. ## Want this model on your inference provider? Ask your inference provider—such as [OpenRouter](https://openrouter.ai/) or another managed inference service—to deploy this exact repository with its multimodal processor and serving plugin. Deploying only the upstream text model will not enable image input. Providers should complete the fresh 0731 GPU validation gates above before offering it as a production endpoint. ## Experimental status, roadmap, and get involved > [!CAUTION] > **Experimental vision checkpoint.** Repository assembly, provenance, hashes, > and source preservation have been verified, but a fresh full GPU loader and > image-generation smoke test has not been run for this 0731 package. This is > not a production-readiness or quality claim. Fine-grained OCR, small-object > and control identification, GUI grounding, hallucination calibration, > concurrency, and broader hardware coverage still need evaluation. Do not use > this model as the sole decision source for safety-critical automation. The > roadmap includes larger and more diverse datasets, higher-resolution OCR/UI > examples, and broader parameter-efficient tuning. Interested in contributing > evaluation or training data, sponsoring compute, or working with us as a > design partner? [Tell us here](https://forms.gle/bNoeJ6cvLYQ4VgKd7).